[SOLVED] Spamassassin per-user bayes training

[garyo]

Does ZCS keep a per-user spam/ham bayes db?

If not, I'd like to make it do that. It gives much better results since it fine-tunes the token set each user is likely to get.

In our spamassassin frontend where we run spamd, I currently execute spamd like this:
/usr/sbin/spamd --create-prefs -u mailuser -x -l --max-children 5 --helper-home-dir --virtual-config-dir=/etc/spamassassin/userprefs/%l -d

This creates and uses per-virtual-user userprefs (including bayes dbs) in /etc/spamassassin/userprefs. I'm not yet sure how to make Zimbra do this. Maybe it invokes SA from amavisd.

[phoenix]

Currently this bug is for spamassassin user prefs: Bug 3870 – per user Spam Assassin score

[garyo]

Got it. Related question: where does Zimbra invoke amavisd? Maybe I can just patch a call to spamd in there.

[garyo]

Found out where amavisd comes in: it's invoked as a filter MTA directly from postfix. So there's no way to hook spamd in there. But, there's really little reason to need to -- amavisd is pretty clever. It only runs SA once per mail even if there's more than one recipient, and it keeps the SA perl mods in memory so it's as fast as spamd. It's just confusing to an experienced postfix/procmail/SA admin like me at first. I think amavisd will have to be enhanced to allow per-user bayes db training, I doubt Zimbra can do it itself with the current architecture.